Nvidia pays Washington to sell to China

Washington has agreed to let Nvidia resume sales of its customised H20 artificial intelligence chips to China, on condition that the company hands 15 per cent of the revenue straight to the US Treasury. It is an odd compromise, somewhere between taxation and a shakedown.

The H20 was designed to fit under export rules written in 2023, then caught in April by a new licensing requirement that amounted to a ban. Nvidia took a $4.5 billion charge on inventory it could not sell and warned that the restrictions would cost it about $8 billion of revenue in its second fiscal quarter. In July the company said it had been assured that licences would be granted. The revenue-sharing condition, reported this week, also covers the MI308 chips that Advanced Micro Devices sells to China.

The controls themselves have moved in stages. The first broad restrictions on advanced AI chips to China came in October 2022, and Nvidia answered with slower versions designed to fit under the limits. The rules were tightened in October 2023 to close those gaps, and the H20 was the product of that second redesign. Each round has pushed Chinese buyers to hedge by ordering domestic chips, and the release of DeepSeek's efficient models in January, trained partly on restricted hardware, suggested that Chinese developers could do more with less than Washington had assumed.

Export controls with a price tag

The deal creates a new tool of trade policy, in which the government charges for its national security export controls. Washington has chosen a toll on overseas sales over an outright ban on advanced chips. Giving up 15 per cent of gross H20 revenue is worth it for Nvidia. It keeps the company in the Chinese market and stops domestic rivals such as Huawei from taking over the country's corporate AI infrastructure.

That trade-off rests on software as much as silicon. Chinese developers built their models on Nvidia's CUDA programming platform, and every quarter they spend porting code to Huawei's Ascend chips makes the switch harder to reverse. Keeping the H20 in Chinese data centres keeps those developers writing for CUDA, and the payment to the Treasury is the price of defending that position.

Beijing may refuse

The strongest objection to the deal's value comes from China rather than Washington. At the end of July the Cyberspace Administration of China summoned Nvidia to explain alleged security risks in the H20, and Chinese officials have every reason to steer big buyers towards domestic suppliers. The chip is also a cut-down version of the Hopper design, a generation behind the Blackwell parts Nvidia sells elsewhere. If Chinese demand does not come back, the licence and the levy both become academic, and much of the written-down inventory stays on the shelf.

There are signs that Beijing is already pushing back. Reports this week said Chinese authorities had discouraged companies from buying the H20, particularly for government-related work. That fits a longer campaign to build a domestic chip industry, backed by state investment funds, and it gives Chinese buyers a patriotic reason to accept slower hardware. Nvidia can pay Washington for permission to sell; it cannot pay Beijing for permission to be bought.

A bad precedent

In accounting terms the 15 per cent charge works like a royalty that cannot be deducted, and it will depress gross margins on Chinese sales for as long as it lasts. It also sets a worrying precedent. Dominant technology companies may now keep their foreign business going only by paying tribute to the government.

The arrangement fits a pattern of the state taking a cut of private deals. The government took a golden share in US Steel as its price for approving Nippon Steel's takeover in June, and in July the Pentagon became the largest shareholder in MP Materials, the rare-earths producer. Lawyers have pointed out that the Constitution forbids taxes or duties on exports, which leaves the legal footing of a levy on overseas chip sales open to challenge.

Shareholders gain in the short run, since 85 cents of every H20 dollar beats nothing. The losers are harder to see: companies without Nvidia's bargaining power, which may find that licences come with steeper terms or none at all, and allies whose own export controls lose force if Washington treats its rules as a revenue source. Huawei loses too if the H20 returns in volume, because the April ban had handed it a captive market for AI accelerators inside China.

The case for the deal is not trivial. A ban did not stop Chinese AI development; it mainly transferred revenue from Nvidia to Huawei and gave Chinese developers a reason to abandon CUDA. A licence with a fee keeps them dependent on American technology that is a generation behind the frontier, and the Treasury collects some money along the way. On that view the fee is a crude but workable compromise between security and commerce, and the precedent is less alarming than it looks.

Security rules for sale

The deeper objection concerns what the payment says about the controls. Export restrictions are justified on the ground that certain technology is too dangerous to sell to a rival military. If a chip can be sold once a fee is paid, the security case for blocking it looks weaker, and China hawks in Congress from both parties have made exactly that criticism. Allies such as the Netherlands and Japan, which restrict sales of chipmaking equipment at Washington's request, may also ask why their companies should forgo Chinese revenue when American ones can buy an exception.

Numbers due this month

Nvidia reports quarterly results on 27 August, and its guidance will show whether it expects H20 revenue to resume and how it will book the payment. No formal rule has yet set out how the money will be collected or under what legal authority, so the terms could still change. Nvidia's arrangement with the Treasury puts a price on export controls, and other chipmakers seeking licences to sell abroad should expect to be handed a similar bill, set just as arbitrarily.